diff --git a/debuggerd/debuggerd_test.cpp b/debuggerd/debuggerd_test.cpp
index 5565e8b..e5af425 100644
--- a/debuggerd/debuggerd_test.cpp
+++ b/debuggerd/debuggerd_test.cpp
@@ -34,7 +34,6 @@
 #include <android/set_abort_message.h>
 #include <bionic/malloc.h>
 #include <bionic/mte.h>
-#include <bionic/mte_kernel.h>
 #include <bionic/reserved_signals.h>
 
 #include <android-base/cmsg.h>
@@ -386,16 +385,6 @@
 
 #if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
 static void SetTagCheckingLevelSync() {
-  int tagged_addr_ctrl = prctl(PR_GET_TAGGED_ADDR_CTRL, 0, 0, 0, 0);
-  if (tagged_addr_ctrl < 0) {
-    abort();
-  }
-
-  tagged_addr_ctrl = (tagged_addr_ctrl & ~PR_MTE_TCF_MASK) | PR_MTE_TCF_SYNC;
-  if (prctl(PR_SET_TAGGED_ADDR_CTRL, tagged_addr_ctrl, 0, 0, 0) != 0) {
-    abort();
-  }
-
   HeapTaggingLevel heap_tagging_level = M_HEAP_TAGGING_LEVEL_SYNC;
   if (!android_mallopt(M_SET_HEAP_TAGGING_LEVEL, &heap_tagging_level, sizeof(heap_tagging_level))) {
     abort();
diff --git a/fs_mgr/libsnapshot/android/snapshot/snapshot.proto b/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
index acb75d0..38c6bf8 100644
--- a/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
+++ b/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
@@ -118,7 +118,7 @@
     Cancelled = 7;
 };
 
-// Next: 5
+// Next: 6
 message SnapshotUpdateStatus {
     UpdateState state = 1;
 
@@ -133,6 +133,9 @@
 
     // Sectors allocated for metadata in all the snapshot devices.
     uint64 metadata_sectors = 4;
+
+    // Whether compression/dm-user was used for any snapshots.
+    bool compression_enabled = 5;
 }
 
 // Next: 4
diff --git a/fs_mgr/libsnapshot/device_info.cpp b/fs_mgr/libsnapshot/device_info.cpp
index 8770255..0e90100 100644
--- a/fs_mgr/libsnapshot/device_info.cpp
+++ b/fs_mgr/libsnapshot/device_info.cpp
@@ -120,9 +120,5 @@
 #endif
 }
 
-std::string DeviceInfo::GetSnapuserdFirstStagePidVar() const {
-    return kSnapuserdFirstStagePidVar;
-}
-
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/device_info.h b/fs_mgr/libsnapshot/device_info.h
index 1f08860..d8d3d91 100644
--- a/fs_mgr/libsnapshot/device_info.h
+++ b/fs_mgr/libsnapshot/device_info.h
@@ -39,7 +39,6 @@
     bool SetBootControlMergeStatus(MergeStatus status) override;
     bool SetSlotAsUnbootable(unsigned int slot) override;
     bool IsRecovery() const override;
-    std::string GetSnapuserdFirstStagePidVar() const override;
 
   private:
     bool EnsureBootHal();
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/mock_device_info.h b/fs_mgr/libsnapshot/include/libsnapshot/mock_device_info.h
index d5c263d..ef9d648 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/mock_device_info.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/mock_device_info.h
@@ -32,7 +32,6 @@
     MOCK_METHOD(bool, SetBootControlMergeStatus, (MergeStatus status), (override));
     MOCK_METHOD(bool, SetSlotAsUnbootable, (unsigned int slot), (override));
     MOCK_METHOD(bool, IsRecovery, (), (const, override));
-    MOCK_METHOD(std::string, GetSnapuserdFirstStagePidVar, (), (const, override));
 };
 
 }  // namespace android::snapshot
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
index 351dce7..f8d3cdc 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
@@ -107,7 +107,6 @@
         virtual bool SetSlotAsUnbootable(unsigned int slot) = 0;
         virtual bool IsRecovery() const = 0;
         virtual bool IsTestDevice() const { return false; }
-        virtual std::string GetSnapuserdFirstStagePidVar() const = 0;
     };
     virtual ~ISnapshotManager() = default;
 
@@ -345,6 +344,18 @@
     bool MapAllSnapshots(const std::chrono::milliseconds& timeout_ms = {}) override;
     bool UnmapAllSnapshots() override;
 
+    // We can't use WaitForFile during first-stage init, because ueventd is not
+    // running and therefore will not automatically create symlinks. Instead,
+    // we let init provide us with the correct function to use to ensure
+    // uevents have been processed and symlink/mknod calls completed.
+    void SetUeventRegenCallback(std::function<bool(const std::string&)> callback) {
+        uevent_regen_callback_ = callback;
+    }
+
+    // If true, compression is enabled for this update. This is used by
+    // first-stage to decide whether to launch snapuserd.
+    bool IsSnapuserdRequired();
+
   private:
     FRIEND_TEST(SnapshotTest, CleanFirstStageMount);
     FRIEND_TEST(SnapshotTest, CreateSnapshot);
@@ -676,6 +687,12 @@
     // Same as above, but for paths only (no major:minor device strings).
     bool GetMappedImageDevicePath(const std::string& device_name, std::string* device_path);
 
+    // Wait for a device to be created by ueventd (eg, its symlink or node to be populated).
+    // This is needed for any code that uses device-mapper path in first-stage init. If
+    // |timeout_ms| is empty or the given device is not a path, WaitForDevice immediately
+    // returns true.
+    bool WaitForDevice(const std::string& device, std::chrono::milliseconds timeout_ms);
+
     std::string gsid_dir_;
     std::string metadata_dir_;
     std::unique_ptr<IDeviceInfo> device_;
@@ -683,6 +700,7 @@
     bool has_local_image_manager_ = false;
     bool use_first_stage_snapuserd_ = false;
     bool in_factory_data_reset_ = false;
+    std::function<bool(const std::string&)> uevent_regen_callback_;
     std::unique_ptr<SnapuserdClient> snapuserd_client_;
 };
 
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
index 0e9ba9e..aa9ba6e 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
@@ -30,7 +30,6 @@
 
 static constexpr uint32_t PACKET_SIZE = 512;
 
-static constexpr char kSnapuserdSocketFirstStage[] = "snapuserd_first_stage";
 static constexpr char kSnapuserdSocket[] = "snapuserd";
 
 static constexpr char kSnapuserdFirstStagePidVar[] = "FIRST_STAGE_SNAPUSERD_PID";
@@ -38,10 +37,6 @@
 // Ensure that the second-stage daemon for snapuserd is running.
 bool EnsureSnapuserdStarted();
 
-// Start the first-stage version of snapuserd, returning its pid. This is used
-// by first-stage init, as well as vts_libsnapshot_test. On failure, -1 is returned.
-pid_t StartFirstStageSnapuserd();
-
 class SnapuserdClient {
   private:
     android::base::unique_fd sockfd_;
@@ -75,6 +70,11 @@
     // Wait for snapuserd to disassociate with a dm-user control device. This
     // must ONLY be called if the control device has already been deleted.
     bool WaitForDeviceDelete(const std::string& control_device);
+
+    // Detach snapuserd. This shuts down the listener socket, and will cause
+    // snapuserd to gracefully exit once all handler threads have terminated.
+    // This should only be used on first-stage instances of snapuserd.
+    bool DetachSnapuserd();
 };
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_server.h b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_server.h
index cadfd71..1491aac 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_server.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_server.h
@@ -41,6 +41,7 @@
     QUERY,
     STOP,
     DELETE,
+    DETACH,
     INVALID,
 };
 
@@ -106,6 +107,7 @@
     bool IsTerminating() { return terminating_; }
 
     void RunThread(std::shared_ptr<DmUserHandler> handler);
+    void JoinAllThreads();
 
     // Find a DmUserHandler within a lock.
     HandlerList::iterator FindHandler(std::lock_guard<std::mutex>* proof_of_lock,
diff --git a/fs_mgr/libsnapshot/include_test/libsnapshot/test_helpers.h b/fs_mgr/libsnapshot/include_test/libsnapshot/test_helpers.h
index 1d1510a..7aef086 100644
--- a/fs_mgr/libsnapshot/include_test/libsnapshot/test_helpers.h
+++ b/fs_mgr/libsnapshot/include_test/libsnapshot/test_helpers.h
@@ -96,7 +96,6 @@
         return true;
     }
     bool IsTestDevice() const override { return true; }
-    std::string GetSnapuserdFirstStagePidVar() const override { return {}; }
 
     bool IsSlotUnbootable(uint32_t slot) { return unbootable_slots_.count(slot) != 0; }
 
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index 10c707a..04c3ccc 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -409,10 +409,12 @@
     if (!dm.CreateDevice(name, table, path, timeout_ms)) {
         return false;
     }
+    if (!WaitForDevice(*path, timeout_ms)) {
+        return false;
+    }
 
     auto control_device = "/dev/dm-user/" + misc_name;
-    if (!android::fs_mgr::WaitForFile(control_device, timeout_ms)) {
-        LOG(ERROR) << "Timed out waiting for dm-user misc device: " << control_device;
+    if (!WaitForDevice(control_device, timeout_ms)) {
         return false;
     }
 
@@ -1340,10 +1342,12 @@
             continue;
         }
 
+        auto misc_name = user_cow_name;
+
         DmTable table;
-        table.Emplace<DmTargetUser>(0, target.spec.length, user_cow_name);
+        table.Emplace<DmTargetUser>(0, target.spec.length, misc_name);
         if (!dm.LoadTableAndActivate(user_cow_name, table)) {
-            LOG(ERROR) << "Unable to swap tables for " << user_cow_name;
+            LOG(ERROR) << "Unable to swap tables for " << misc_name;
             continue;
         }
 
@@ -1369,14 +1373,13 @@
         }
 
         // Wait for ueventd to acknowledge and create the control device node.
-        std::string control_device = "/dev/dm-user/" + user_cow_name;
-        if (!android::fs_mgr::WaitForFile(control_device, 10s)) {
-            LOG(ERROR) << "Could not find control device: " << control_device;
+        std::string control_device = "/dev/dm-user/" + misc_name;
+        if (!WaitForDevice(control_device, 10s)) {
             continue;
         }
 
         uint64_t base_sectors =
-                snapuserd_client_->InitDmUserCow(user_cow_name, cow_image_device, backing_device);
+                snapuserd_client_->InitDmUserCow(misc_name, cow_image_device, backing_device);
         if (base_sectors == 0) {
             // Unrecoverable as metadata reads from cow device failed
             LOG(FATAL) << "Failed to retrieve base_sectors from Snapuserd";
@@ -1385,7 +1388,7 @@
 
         CHECK(base_sectors == target.spec.length);
 
-        if (!snapuserd_client_->AttachDmUser(user_cow_name)) {
+        if (!snapuserd_client_->AttachDmUser(misc_name)) {
             // This error is unrecoverable. We cannot proceed because reads to
             // the underlying device will fail.
             LOG(FATAL) << "Could not initialize snapuserd for " << user_cow_name;
@@ -1399,24 +1402,6 @@
         LOG(ERROR) << "Could not transition all snapuserd consumers.";
         return false;
     }
-
-    std::string pid_var = device_->GetSnapuserdFirstStagePidVar();
-    if (pid_var.empty()) {
-        return true;
-    }
-
-    int pid;
-    const char* pid_str = getenv(pid_var.c_str());
-    if (pid_str && android::base::ParseInt(pid_str, &pid)) {
-        if (kill(pid, SIGTERM) < 0 && errno != ESRCH) {
-            LOG(ERROR) << "kill snapuserd failed";
-            return false;
-        }
-    } else {
-        LOG(ERROR) << "Could not find or parse " << kSnapuserdFirstStagePidVar
-                   << " for snapuserd pid";
-        return false;
-    }
     return true;
 }
 
@@ -1761,15 +1746,6 @@
         }
     }
 
-    if (use_first_stage_snapuserd_) {
-        // Remove the first-stage socket as a precaution, there is no need to
-        // access the daemon anymore and we'll be killing it once second-stage
-        // is running.
-        auto socket = ANDROID_SOCKET_DIR + "/"s + kSnapuserdSocketFirstStage;
-        snapuserd_client_ = nullptr;
-        unlink(socket.c_str());
-    }
-
     LOG(INFO) << "Created logical partitions with snapshot.";
     return true;
 }
@@ -1924,13 +1900,20 @@
     if (live_snapshot_status->compression_enabled()) {
         auto name = GetDmUserCowName(params.GetPartitionName());
 
-        // :TODO: need to force init to process uevents for these in first-stage.
         std::string cow_path;
         if (!GetMappedImageDevicePath(cow_name, &cow_path)) {
             LOG(ERROR) << "Could not determine path for: " << cow_name;
             return false;
         }
 
+        // Ensure both |base_path| and |cow_path| are created, for snapuserd.
+        if (!WaitForDevice(base_path, remaining_time)) {
+            return false;
+        }
+        if (!WaitForDevice(cow_path, remaining_time)) {
+            return false;
+        }
+
         std::string new_cow_device;
         if (!MapDmUserCow(lock, name, cow_path, base_path, remaining_time, &new_cow_device)) {
             LOG(ERROR) << "Could not map dm-user device for partition "
@@ -2070,7 +2053,7 @@
         if (!EnsureSnapuserdConnected()) {
             return false;
         }
-        if (!dm.DeleteDevice(dm_user_name)) {
+        if (!dm.DeleteDeviceIfExists(dm_user_name)) {
             LOG(ERROR) << "Cannot unmap " << dm_user_name;
             return false;
         }
@@ -2273,6 +2256,7 @@
 bool SnapshotManager::WriteUpdateState(LockedFile* lock, UpdateState state) {
     SnapshotUpdateStatus status = {};
     status.set_state(state);
+    status.set_compression_enabled(IsCompressionEnabled());
     return WriteSnapshotUpdateStatus(lock, status);
 }
 
@@ -2408,28 +2392,11 @@
         return true;
     }
 
-    std::string socket;
-    if (use_first_stage_snapuserd_) {
-        auto pid = StartFirstStageSnapuserd();
-        if (pid < 0) {
-            LOG(ERROR) << "Failed to start snapuserd";
-            return false;
-        }
-
-        auto pid_str = std::to_string(static_cast<int>(pid));
-        if (setenv(kSnapuserdFirstStagePidVar, pid_str.c_str(), 1) < 0) {
-            PLOG(ERROR) << "setenv failed storing the snapuserd pid";
-        }
-
-        socket = kSnapuserdSocketFirstStage;
-    } else {
-        if (!EnsureSnapuserdStarted()) {
-            return false;
-        }
-        socket = kSnapuserdSocket;
+    if (!use_first_stage_snapuserd_ && !EnsureSnapuserdStarted()) {
+        return false;
     }
 
-    snapuserd_client_ = SnapuserdClient::Connect(socket, 10s);
+    snapuserd_client_ = SnapuserdClient::Connect(kSnapuserdSocket, 10s);
     if (!snapuserd_client_) {
         LOG(ERROR) << "Unable to connect to snapuserd";
         return false;
@@ -2845,7 +2812,7 @@
         return nullptr;
     }
 
-    if (IsCompressionEnabled()) {
+    if (status.compression_enabled()) {
         return OpenCompressedSnapshotWriter(lock.get(), source_device, params.GetPartitionName(),
                                             status, paths);
     }
@@ -3302,5 +3269,47 @@
     return true;
 }
 
+bool SnapshotManager::WaitForDevice(const std::string& device,
+                                    std::chrono::milliseconds timeout_ms) {
+    if (!android::base::StartsWith(device, "/")) {
+        return true;
+    }
+
+    // In first-stage init, we rely on init setting a callback which can
+    // regenerate uevents and populate /dev for us.
+    if (uevent_regen_callback_) {
+        if (!uevent_regen_callback_(device)) {
+            LOG(ERROR) << "Failed to find device after regenerating uevents: " << device;
+            return false;
+        }
+        return true;
+    }
+
+    // Otherwise, the only kind of device we need to wait for is a dm-user
+    // misc device. Normal calls to DeviceMapper::CreateDevice() guarantee
+    // the path has been created.
+    if (!android::base::StartsWith(device, "/dev/dm-user/")) {
+        return true;
+    }
+
+    if (timeout_ms.count() == 0) {
+        LOG(ERROR) << "No timeout was specified to wait for device: " << device;
+        return false;
+    }
+    if (!android::fs_mgr::WaitForFile(device, timeout_ms)) {
+        LOG(ERROR) << "Timed out waiting for device to appear: " << device;
+        return false;
+    }
+    return true;
+}
+
+bool SnapshotManager::IsSnapuserdRequired() {
+    auto lock = LockExclusive();
+    if (!lock) return false;
+
+    auto status = ReadSnapshotUpdateStatus(lock.get());
+    return status.state() != UpdateState::None && status.compression_enabled();
+}
+
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/snapshot_fuzz_utils.h b/fs_mgr/libsnapshot/snapshot_fuzz_utils.h
index 092ee0b..5319e69 100644
--- a/fs_mgr/libsnapshot/snapshot_fuzz_utils.h
+++ b/fs_mgr/libsnapshot/snapshot_fuzz_utils.h
@@ -124,7 +124,6 @@
         return data_->allow_set_slot_as_unbootable();
     }
     bool IsRecovery() const override { return data_->is_recovery(); }
-    std::string GetSnapuserdFirstStagePidVar() const override { return {}; }
 
     void SwitchSlot() { switched_slot_ = !switched_slot_; }
 
diff --git a/fs_mgr/libsnapshot/snapshot_test.cpp b/fs_mgr/libsnapshot/snapshot_test.cpp
index 74558ab..c6a1786 100644
--- a/fs_mgr/libsnapshot/snapshot_test.cpp
+++ b/fs_mgr/libsnapshot/snapshot_test.cpp
@@ -1760,9 +1760,6 @@
         GTEST_SKIP() << "Skipping Virtual A/B Compression test";
     }
 
-    AutoKill auto_kill(StartFirstStageSnapuserd());
-    ASSERT_TRUE(auto_kill.valid());
-
     // Ensure a connection to the second-stage daemon, but use the first-stage
     // code paths thereafter.
     ASSERT_TRUE(sm->EnsureSnapuserdConnected());
diff --git a/fs_mgr/libsnapshot/snapuserd.cpp b/fs_mgr/libsnapshot/snapuserd.cpp
index 954699c..49e6c3d 100644
--- a/fs_mgr/libsnapshot/snapuserd.cpp
+++ b/fs_mgr/libsnapshot/snapuserd.cpp
@@ -28,6 +28,9 @@
 using namespace android::dm;
 using android::base::unique_fd;
 
+#define SNAP_LOG(level) LOG(level) << misc_name_ << ": "
+#define SNAP_PLOG(level) PLOG(level) << misc_name_ << ": "
+
 static constexpr size_t PAYLOAD_SIZE = (1UL << 16);
 
 static_assert(PAYLOAD_SIZE >= BLOCK_SIZE);
@@ -94,7 +97,7 @@
 // it will be de-compressed.
 bool Snapuserd::ProcessReplaceOp(const CowOperation* cow_op) {
     if (!reader_->ReadData(*cow_op, &bufsink_)) {
-        LOG(ERROR) << "ReadData failed for chunk: " << cow_op->new_block;
+        SNAP_LOG(ERROR) << "ReadData failed for chunk: " << cow_op->new_block;
         return false;
     }
 
@@ -111,7 +114,7 @@
     // if the successive blocks are contiguous.
     if (!android::base::ReadFullyAtOffset(backing_store_fd_, buffer, BLOCK_SIZE,
                                           cow_op->source * BLOCK_SIZE)) {
-        LOG(ERROR) << "Copy-op failed. Read from backing store at: " << cow_op->source;
+        SNAP_LOG(ERROR) << "Copy-op failed. Read from backing store at: " << cow_op->source;
         return false;
     }
 
@@ -180,14 +183,14 @@
             }
 
             default: {
-                LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
+                SNAP_LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
                 ret = false;
                 break;
             }
         }
 
         if (!ret) {
-            LOG(ERROR) << "ReadData failed for operation: " << cow_op->type;
+            SNAP_LOG(ERROR) << "ReadData failed for operation: " << cow_op->type;
             return false;
         }
 
@@ -322,7 +325,7 @@
 
     CHECK(!(*unmerged_exceptions == exceptions_per_area_));
 
-    LOG(DEBUG) << "Unmerged_Exceptions: " << *unmerged_exceptions << " Offset: " << offset;
+    SNAP_LOG(DEBUG) << "Unmerged_Exceptions: " << *unmerged_exceptions << " Offset: " << offset;
     return offset;
 }
 
@@ -331,7 +334,7 @@
     int merged_ops_cur_iter = 0;
 
     // Find the operations which are merged in this cycle.
-    while ((unmerged_exceptions + merged_ops_cur_iter) <= exceptions_per_area_) {
+    while ((unmerged_exceptions + merged_ops_cur_iter) < exceptions_per_area_) {
         struct disk_exception* merged_de =
                 reinterpret_cast<struct disk_exception*>((char*)merged_buffer + offset);
         struct disk_exception* cow_de =
@@ -355,11 +358,11 @@
             CHECK(cow_de->new_chunk == 0);
             break;
         } else {
-            LOG(ERROR) << "Error in merge operation. Found invalid metadata";
-            LOG(ERROR) << "merged_de-old-chunk: " << merged_de->old_chunk;
-            LOG(ERROR) << "merged_de-new-chunk: " << merged_de->new_chunk;
-            LOG(ERROR) << "cow_de-old-chunk: " << cow_de->old_chunk;
-            LOG(ERROR) << "cow_de-new-chunk: " << cow_de->new_chunk;
+            SNAP_LOG(ERROR) << "Error in merge operation. Found invalid metadata";
+            SNAP_LOG(ERROR) << "merged_de-old-chunk: " << merged_de->old_chunk;
+            SNAP_LOG(ERROR) << "merged_de-new-chunk: " << merged_de->new_chunk;
+            SNAP_LOG(ERROR) << "cow_de-old-chunk: " << cow_de->old_chunk;
+            SNAP_LOG(ERROR) << "cow_de-new-chunk: " << cow_de->new_chunk;
             return -1;
         }
     }
@@ -384,19 +387,19 @@
 
         if (!(cow_op->type == kCowReplaceOp || cow_op->type == kCowZeroOp ||
               cow_op->type == kCowCopyOp)) {
-            LOG(ERROR) << "Unknown operation-type found during merge: " << cow_op->type;
+            SNAP_LOG(ERROR) << "Unknown operation-type found during merge: " << cow_op->type;
             return false;
         }
 
         merged_ops_cur_iter -= 1;
-        LOG(DEBUG) << "Merge op found of type " << cow_op->type
-                   << "Pending-merge-ops: " << merged_ops_cur_iter;
+        SNAP_LOG(DEBUG) << "Merge op found of type " << cow_op->type
+                        << "Pending-merge-ops: " << merged_ops_cur_iter;
         cowop_iter_->Next();
     }
 
     if (cowop_iter_->Done()) {
         CHECK(merged_ops_cur_iter == 0);
-        LOG(DEBUG) << "All cow operations merged successfully in this cycle";
+        SNAP_LOG(DEBUG) << "All cow operations merged successfully in this cycle";
     }
 
     return true;
@@ -407,14 +410,15 @@
     CowHeader header;
 
     if (!reader_->GetHeader(&header)) {
-        LOG(ERROR) << "Failed to get header";
+        SNAP_LOG(ERROR) << "Failed to get header";
         return false;
     }
 
     // ChunkID to vector index
     lldiv_t divresult = lldiv(chunk, stride);
     CHECK(divresult.quot < vec_.size());
-    LOG(DEBUG) << "ProcessMergeComplete: chunk: " << chunk << " Metadata-Index: " << divresult.quot;
+    SNAP_LOG(DEBUG) << "ProcessMergeComplete: chunk: " << chunk
+                    << " Metadata-Index: " << divresult.quot;
 
     int unmerged_exceptions = 0;
     loff_t offset = GetMergeStartOffset(buffer, vec_[divresult.quot].get(), &unmerged_exceptions);
@@ -429,11 +433,11 @@
     header.num_merge_ops += merged_ops_cur_iter;
     reader_->UpdateMergeProgress(merged_ops_cur_iter);
     if (!writer_->CommitMerge(merged_ops_cur_iter)) {
-        LOG(ERROR) << "CommitMerge failed...";
+        SNAP_LOG(ERROR) << "CommitMerge failed...";
         return false;
     }
 
-    LOG(DEBUG) << "Merge success";
+    SNAP_LOG(DEBUG) << "Merge success";
     return true;
 }
 
@@ -513,21 +517,21 @@
     bool prev_copy_op = false;
     bool metadata_found = false;
 
-    LOG(DEBUG) << "ReadMetadata Start...";
+    SNAP_LOG(DEBUG) << "ReadMetadata Start...";
 
     if (!reader_->Parse(cow_fd_)) {
-        LOG(ERROR) << "Failed to parse";
+        SNAP_LOG(ERROR) << "Failed to parse";
         return false;
     }
 
     if (!reader_->GetHeader(&header)) {
-        LOG(ERROR) << "Failed to get header";
+        SNAP_LOG(ERROR) << "Failed to get header";
         return false;
     }
 
     CHECK(header.block_size == BLOCK_SIZE);
 
-    LOG(DEBUG) << "Merge-ops: " << header.num_merge_ops;
+    SNAP_LOG(DEBUG) << "Merge-ops: " << header.num_merge_ops;
 
     writer_ = std::make_unique<CowWriter>(options);
     writer_->InitializeMerge(cow_fd_.get(), &header);
@@ -563,7 +567,7 @@
 
         if (!(cow_op->type == kCowReplaceOp || cow_op->type == kCowZeroOp ||
               cow_op->type == kCowCopyOp)) {
-            LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
+            SNAP_LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
             return false;
         }
 
@@ -578,7 +582,7 @@
         de->old_chunk = cow_op->new_block;
         de->new_chunk = next_free;
 
-        LOG(DEBUG) << "Old-chunk: " << de->old_chunk << "New-chunk: " << de->new_chunk;
+        SNAP_LOG(DEBUG) << "Old-chunk: " << de->old_chunk << "New-chunk: " << de->new_chunk;
 
         // Store operation pointer.
         chunk_map_[next_free] = cow_op;
@@ -602,7 +606,7 @@
 
             if (cowop_riter_->Done()) {
                 vec_.push_back(std::move(de_ptr));
-                LOG(DEBUG) << "ReadMetadata() completed; Number of Areas: " << vec_.size();
+                SNAP_LOG(DEBUG) << "ReadMetadata() completed; Number of Areas: " << vec_.size();
             }
         }
 
@@ -614,12 +618,12 @@
     // is aware that merge is completed.
     if (num_ops || !metadata_found) {
         vec_.push_back(std::move(de_ptr));
-        LOG(DEBUG) << "ReadMetadata() completed. Partially filled area num_ops: " << num_ops
-                   << "Areas : " << vec_.size();
+        SNAP_LOG(DEBUG) << "ReadMetadata() completed. Partially filled area num_ops: " << num_ops
+                        << "Areas : " << vec_.size();
     }
 
-    LOG(DEBUG) << "ReadMetadata() completed. chunk_id: " << next_free
-               << "Num Sector: " << ChunkToSector(next_free);
+    SNAP_LOG(DEBUG) << "ReadMetadata() completed. chunk_id: " << next_free
+                    << "Num Sector: " << ChunkToSector(next_free);
 
     // Initialize the iterator for merging
     cowop_iter_ = reader_->GetOpIter();
@@ -643,7 +647,7 @@
 // us the sector number for which IO is issued by dm-snapshot device
 bool Snapuserd::ReadDmUserHeader() {
     if (!android::base::ReadFully(ctrl_fd_, bufsink_.GetBufPtr(), sizeof(struct dm_user_header))) {
-        PLOG(ERROR) << "ReadDmUserHeader failed";
+        SNAP_PLOG(ERROR) << "Control-read failed";
         return false;
     }
 
@@ -654,7 +658,7 @@
 bool Snapuserd::WriteDmUserPayload(size_t size) {
     if (!android::base::WriteFully(ctrl_fd_, bufsink_.GetBufPtr(),
                                    sizeof(struct dm_user_header) + size)) {
-        PLOG(ERROR) << "Write to dm-user failed";
+        SNAP_PLOG(ERROR) << "Write to dm-user failed";
         return false;
     }
 
@@ -663,7 +667,7 @@
 
 bool Snapuserd::ReadDmUserPayload(void* buffer, size_t size) {
     if (!android::base::ReadFully(ctrl_fd_, buffer, size)) {
-        PLOG(ERROR) << "ReadDmUserPayload failed";
+        SNAP_PLOG(ERROR) << "ReadDmUserPayload failed";
         return false;
     }
 
@@ -673,7 +677,7 @@
 bool Snapuserd::InitCowDevice() {
     cow_fd_.reset(open(cow_device_.c_str(), O_RDWR));
     if (cow_fd_ < 0) {
-        PLOG(ERROR) << "Open Failed: " << cow_device_;
+        SNAP_PLOG(ERROR) << "Open Failed: " << cow_device_;
         return false;
     }
 
@@ -690,13 +694,13 @@
 bool Snapuserd::InitBackingAndControlDevice() {
     backing_store_fd_.reset(open(backing_store_device_.c_str(), O_RDONLY));
     if (backing_store_fd_ < 0) {
-        PLOG(ERROR) << "Open Failed: " << backing_store_device_;
+        SNAP_PLOG(ERROR) << "Open Failed: " << backing_store_device_;
         return false;
     }
 
     ctrl_fd_.reset(open(control_device_.c_str(), O_RDWR));
     if (ctrl_fd_ < 0) {
-        PLOG(ERROR) << "Unable to open " << control_device_;
+        SNAP_PLOG(ERROR) << "Unable to open " << control_device_;
         return false;
     }
 
@@ -709,15 +713,15 @@
     bufsink_.Clear();
 
     if (!ReadDmUserHeader()) {
-        LOG(ERROR) << "ReadDmUserHeader failed";
+        SNAP_LOG(ERROR) << "ReadDmUserHeader failed";
         return false;
     }
 
-    LOG(DEBUG) << "msg->seq: " << std::hex << header->seq;
-    LOG(DEBUG) << "msg->type: " << std::hex << header->type;
-    LOG(DEBUG) << "msg->flags: " << std::hex << header->flags;
-    LOG(DEBUG) << "msg->sector: " << std::hex << header->sector;
-    LOG(DEBUG) << "msg->len: " << std::hex << header->len;
+    SNAP_LOG(DEBUG) << "msg->seq: " << std::hex << header->seq;
+    SNAP_LOG(DEBUG) << "msg->type: " << std::hex << header->type;
+    SNAP_LOG(DEBUG) << "msg->flags: " << std::hex << header->flags;
+    SNAP_LOG(DEBUG) << "msg->sector: " << std::hex << header->sector;
+    SNAP_LOG(DEBUG) << "msg->len: " << std::hex << header->len;
 
     switch (header->type) {
         case DM_USER_REQ_MAP_READ: {
@@ -736,7 +740,7 @@
                     CHECK(metadata_read_done_ == true);
                     CHECK(read_size == BLOCK_SIZE);
                     ConstructKernelCowHeader();
-                    LOG(DEBUG) << "Kernel header constructed";
+                    SNAP_LOG(DEBUG) << "Kernel header constructed";
                 } else {
                     // Convert the sector number to a chunk ID.
                     //
@@ -747,22 +751,22 @@
 
                     if (chunk_map_.find(chunk) == chunk_map_.end()) {
                         if (!ReadDiskExceptions(chunk, read_size)) {
-                            LOG(ERROR) << "ReadDiskExceptions failed for chunk id: " << chunk
-                                       << "Sector: " << header->sector;
+                            SNAP_LOG(ERROR) << "ReadDiskExceptions failed for chunk id: " << chunk
+                                            << "Sector: " << header->sector;
                             header->type = DM_USER_RESP_ERROR;
                         } else {
-                            LOG(DEBUG) << "ReadDiskExceptions success for chunk id: " << chunk
-                                       << "Sector: " << header->sector;
+                            SNAP_LOG(DEBUG) << "ReadDiskExceptions success for chunk id: " << chunk
+                                            << "Sector: " << header->sector;
                         }
                     } else {
                         chunk_t num_chunks_read = (offset >> BLOCK_SHIFT);
                         if (!ReadData(chunk + num_chunks_read, read_size)) {
-                            LOG(ERROR) << "ReadData failed for chunk id: " << chunk
-                                       << "Sector: " << header->sector;
+                            SNAP_LOG(ERROR) << "ReadData failed for chunk id: " << chunk
+                                            << "Sector: " << header->sector;
                             header->type = DM_USER_RESP_ERROR;
                         } else {
-                            LOG(DEBUG) << "ReadData success for chunk id: " << chunk
-                                       << "Sector: " << header->sector;
+                            SNAP_LOG(DEBUG) << "ReadData success for chunk id: " << chunk
+                                            << "Sector: " << header->sector;
                         }
                     }
                 }
@@ -817,18 +821,18 @@
             header->type = DM_USER_RESP_SUCCESS;
 
             if (!ReadDmUserPayload(buffer, read_size)) {
-                LOG(ERROR) << "ReadDmUserPayload failed for chunk id: " << chunk
-                           << "Sector: " << header->sector;
+                SNAP_LOG(ERROR) << "ReadDmUserPayload failed for chunk id: " << chunk
+                                << "Sector: " << header->sector;
                 header->type = DM_USER_RESP_ERROR;
             }
 
             if (header->type == DM_USER_RESP_SUCCESS && !ProcessMergeComplete(chunk, buffer)) {
-                LOG(ERROR) << "ProcessMergeComplete failed for chunk id: " << chunk
-                           << "Sector: " << header->sector;
+                SNAP_LOG(ERROR) << "ProcessMergeComplete failed for chunk id: " << chunk
+                                << "Sector: " << header->sector;
                 header->type = DM_USER_RESP_ERROR;
             } else {
-                LOG(DEBUG) << "ProcessMergeComplete success for chunk id: " << chunk
-                           << "Sector: " << header->sector;
+                SNAP_LOG(DEBUG) << "ProcessMergeComplete success for chunk id: " << chunk
+                                << "Sector: " << header->sector;
             }
 
             if (!WriteDmUserPayload(0)) {
diff --git a/fs_mgr/libsnapshot/snapuserd_client.cpp b/fs_mgr/libsnapshot/snapuserd_client.cpp
index a5d2061..7282bff 100644
--- a/fs_mgr/libsnapshot/snapuserd_client.cpp
+++ b/fs_mgr/libsnapshot/snapuserd_client.cpp
@@ -54,25 +54,6 @@
     return true;
 }
 
-pid_t StartFirstStageSnapuserd() {
-    pid_t pid = fork();
-    if (pid < 0) {
-        PLOG(ERROR) << "fork failed";
-        return pid;
-    }
-    if (pid != 0) {
-        return pid;
-    }
-
-    std::string arg0 = "/system/bin/snapuserd";
-    std::string arg1 = kSnapuserdSocketFirstStage;
-    char* const argv[] = {arg0.data(), arg1.data(), nullptr};
-    if (execv(arg0.c_str(), argv) < 0) {
-        PLOG(FATAL) << "execv failed";
-    }
-    return pid;
-}
-
 SnapuserdClient::SnapuserdClient(android::base::unique_fd&& sockfd) : sockfd_(std::move(sockfd)) {}
 
 static inline bool IsRetryErrno() {
@@ -229,5 +210,13 @@
     return num_sectors;
 }
 
+bool SnapuserdClient::DetachSnapuserd() {
+    if (!Sendmsg("detach")) {
+        LOG(ERROR) << "Failed to detach snapuserd.";
+        return false;
+    }
+    return true;
+}
+
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/snapuserd_server.cpp b/fs_mgr/libsnapshot/snapuserd_server.cpp
index 9d57ab0..7a5cead 100644
--- a/fs_mgr/libsnapshot/snapuserd_server.cpp
+++ b/fs_mgr/libsnapshot/snapuserd_server.cpp
@@ -38,6 +38,7 @@
     if (input == "stop") return DaemonOperations::STOP;
     if (input == "query") return DaemonOperations::QUERY;
     if (input == "delete") return DaemonOperations::DELETE;
+    if (input == "detach") return DaemonOperations::DETACH;
 
     return DaemonOperations::INVALID;
 }
@@ -72,19 +73,7 @@
 
 void SnapuserdServer::ShutdownThreads() {
     StopThreads();
-
-    // Acquire the thread list within the lock.
-    std::vector<std::shared_ptr<DmUserHandler>> dm_users;
-    {
-        std::lock_guard<std::mutex> guard(lock_);
-        dm_users = std::move(dm_users_);
-    }
-
-    for (auto& client : dm_users) {
-        auto& th = client->thread();
-
-        if (th.joinable()) th.join();
-    }
+    JoinAllThreads();
 }
 
 const std::string& DmUserHandler::GetMiscName() const {
@@ -214,6 +203,10 @@
             }
             return Sendmsg(fd, "success");
         }
+        case DaemonOperations::DETACH: {
+            terminating_ = true;
+            return Sendmsg(fd, "success");
+        }
         default: {
             LOG(ERROR) << "Received unknown message type from client";
             Sendmsg(fd, "fail");
@@ -234,7 +227,7 @@
 
     LOG(INFO) << "Exiting thread for handler: " << handler->GetMiscName();
 
-    // If the main thread called /emoveHandler, the handler was already removed
+    // If the main thread called RemoveHandler, the handler was already removed
     // from within the lock, and calling RemoveHandler again has no effect.
     RemoveHandler(handler->GetMiscName(), false);
 }
@@ -286,9 +279,26 @@
             }
         }
     }
+
+    JoinAllThreads();
     return true;
 }
 
+void SnapuserdServer::JoinAllThreads() {
+    // Acquire the thread list within the lock.
+    std::vector<std::shared_ptr<DmUserHandler>> dm_users;
+    {
+        std::lock_guard<std::mutex> guard(lock_);
+        dm_users = std::move(dm_users_);
+    }
+
+    for (auto& client : dm_users) {
+        auto& th = client->thread();
+
+        if (th.joinable()) th.join();
+    }
+}
+
 void SnapuserdServer::AddWatchedFd(android::base::borrowed_fd fd) {
     struct pollfd p = {};
     p.fd = fd.get();
diff --git a/fs_mgr/tools/Android.bp b/fs_mgr/tools/Android.bp
index 4d4aae4..d6ccc4b 100644
--- a/fs_mgr/tools/Android.bp
+++ b/fs_mgr/tools/Android.bp
@@ -29,3 +29,15 @@
 
     cflags: ["-Werror"],
 }
+
+cc_binary {
+    name: "dmuserd",
+    srcs: ["dmuserd.cpp"],
+
+    shared_libs: [
+        "libbase",
+        "liblog",
+    ],
+
+    cflags: ["-Werror"],
+}
diff --git a/fs_mgr/tools/dmuserd.cpp b/fs_mgr/tools/dmuserd.cpp
new file mode 100644
index 0000000..e50a4a2
--- /dev/null
+++ b/fs_mgr/tools/dmuserd.cpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: Apache-2.0
+
+#define _LARGEFILE64_SOURCE
+
+#include <errno.h>
+#include <fcntl.h>
+#include <getopt.h>
+#include <poll.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/prctl.h>
+#include <unistd.h>
+#include <iostream>
+
+#define SECTOR_SIZE ((__u64)512)
+#define BUFFER_BYTES 4096
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+/* This should be replaced with linux/dm-user.h. */
+#ifndef _LINUX_DM_USER_H
+#define _LINUX_DM_USER_H
+
+#include <linux/types.h>
+
+#define DM_USER_REQ_MAP_READ 0
+#define DM_USER_REQ_MAP_WRITE 1
+#define DM_USER_REQ_MAP_FLUSH 2
+#define DM_USER_REQ_MAP_DISCARD 3
+#define DM_USER_REQ_MAP_SECURE_ERASE 4
+#define DM_USER_REQ_MAP_WRITE_SAME 5
+#define DM_USER_REQ_MAP_WRITE_ZEROES 6
+#define DM_USER_REQ_MAP_ZONE_OPEN 7
+#define DM_USER_REQ_MAP_ZONE_CLOSE 8
+#define DM_USER_REQ_MAP_ZONE_FINISH 9
+#define DM_USER_REQ_MAP_ZONE_APPEND 10
+#define DM_USER_REQ_MAP_ZONE_RESET 11
+#define DM_USER_REQ_MAP_ZONE_RESET_ALL 12
+
+#define DM_USER_REQ_MAP_FLAG_FAILFAST_DEV 0x00001
+#define DM_USER_REQ_MAP_FLAG_FAILFAST_TRANSPORT 0x00002
+#define DM_USER_REQ_MAP_FLAG_FAILFAST_DRIVER 0x00004
+#define DM_USER_REQ_MAP_FLAG_SYNC 0x00008
+#define DM_USER_REQ_MAP_FLAG_META 0x00010
+#define DM_USER_REQ_MAP_FLAG_PRIO 0x00020
+#define DM_USER_REQ_MAP_FLAG_NOMERGE 0x00040
+#define DM_USER_REQ_MAP_FLAG_IDLE 0x00080
+#define DM_USER_REQ_MAP_FLAG_INTEGRITY 0x00100
+#define DM_USER_REQ_MAP_FLAG_FUA 0x00200
+#define DM_USER_REQ_MAP_FLAG_PREFLUSH 0x00400
+#define DM_USER_REQ_MAP_FLAG_RAHEAD 0x00800
+#define DM_USER_REQ_MAP_FLAG_BACKGROUND 0x01000
+#define DM_USER_REQ_MAP_FLAG_NOWAIT 0x02000
+#define DM_USER_REQ_MAP_FLAG_CGROUP_PUNT 0x04000
+#define DM_USER_REQ_MAP_FLAG_NOUNMAP 0x08000
+#define DM_USER_REQ_MAP_FLAG_HIPRI 0x10000
+#define DM_USER_REQ_MAP_FLAG_DRV 0x20000
+#define DM_USER_REQ_MAP_FLAG_SWAP 0x40000
+
+#define DM_USER_RESP_SUCCESS 0
+#define DM_USER_RESP_ERROR 1
+#define DM_USER_RESP_UNSUPPORTED 2
+
+struct dm_user_message {
+    __u64 seq;
+    __u64 type;
+    __u64 flags;
+    __u64 sector;
+    __u64 len;
+    __u8 buf[];
+};
+
+#endif
+
+static bool verbose = false;
+
+ssize_t write_all(int fd, void* buf, size_t len) {
+    char* buf_c = (char*)buf;
+    ssize_t total = 0;
+    ssize_t once;
+
+    while (total < len) {
+        once = write(fd, buf_c + total, len - total);
+        if (once < 0) return once;
+        if (once == 0) {
+            errno = ENOSPC;
+            return 0;
+        }
+        total += once;
+    }
+
+    return total;
+}
+
+ssize_t read_all(int fd, void* buf, size_t len) {
+    char* buf_c = (char*)buf;
+    ssize_t total = 0;
+    ssize_t once;
+
+    while (total < len) {
+        once = read(fd, buf_c + total, len - total);
+        if (once < 0) return once;
+        if (once == 0) {
+            errno = ENOSPC;
+            return 0;
+        }
+        total += once;
+    }
+
+    return total;
+}
+
+int not_splice(int from, int to, __u64 count) {
+    while (count > 0) {
+        char buf[BUFFER_BYTES];
+        __u64 max = count > BUFFER_BYTES ? BUFFER_BYTES : count;
+
+        if (read_all(from, buf, max) <= 0) {
+            perror("Unable to read");
+            return -EIO;
+        }
+
+        if (write_all(to, buf, max) <= 0) {
+            perror("Unable to write");
+            return -EIO;
+        }
+
+        count -= max;
+    }
+
+    return 0;
+}
+
+int simple_daemon(char* control_path, char* backing_path) {
+    int control_fd = open(control_path, O_RDWR);
+    if (control_fd < 0) {
+        fprintf(stderr, "Unable to open control device %s\n", control_path);
+        return -1;
+    }
+
+    int backing_fd = open(backing_path, O_RDWR);
+    if (backing_fd < 0) {
+        fprintf(stderr, "Unable to open backing device %s\n", backing_path);
+        return -1;
+    }
+
+    while (1) {
+        struct dm_user_message msg;
+        char* base;
+        __u64 type;
+
+        if (verbose) std::cerr << "dmuserd: Waiting for message...\n";
+
+        if (read_all(control_fd, &msg, sizeof(msg)) < 0) {
+            if (errno == ENOTBLK) return 0;
+
+            perror("unable to read msg");
+            return -1;
+        }
+
+        if (verbose) {
+            std::string type;
+            switch (msg.type) {
+                case DM_USER_REQ_MAP_WRITE:
+                    type = "write";
+                    break;
+                case DM_USER_REQ_MAP_READ:
+                    type = "read";
+                    break;
+                case DM_USER_REQ_MAP_FLUSH:
+                    type = "flush";
+                    break;
+                default:
+                    /*
+                     * FIXME: Can't I do "whatever"s here rather that
+                     * std::string("whatever")?
+                     */
+                    type = std::string("(unknown, id=") + std::to_string(msg.type) + ")";
+                    break;
+            }
+
+            std::string flags;
+            if (msg.flags & DM_USER_REQ_MAP_FLAG_SYNC) {
+                if (!flags.empty()) flags += "|";
+                flags += "S";
+            }
+            if (msg.flags & DM_USER_REQ_MAP_FLAG_META) {
+                if (!flags.empty()) flags += "|";
+                flags += "M";
+            }
+            if (msg.flags & DM_USER_REQ_MAP_FLAG_FUA) {
+                if (!flags.empty()) flags += "|";
+                flags += "FUA";
+            }
+            if (msg.flags & DM_USER_REQ_MAP_FLAG_PREFLUSH) {
+                if (!flags.empty()) flags += "|";
+                flags += "F";
+            }
+
+            std::cerr << "dmuserd: Got " << type << " request " << flags << " for sector "
+                      << std::to_string(msg.sector) << " with length " << std::to_string(msg.len)
+                      << "\n";
+        }
+
+        type = msg.type;
+        switch (type) {
+            case DM_USER_REQ_MAP_READ:
+                msg.type = DM_USER_RESP_SUCCESS;
+                break;
+            case DM_USER_REQ_MAP_WRITE:
+                if (msg.flags & DM_USER_REQ_MAP_FLAG_PREFLUSH ||
+                    msg.flags & DM_USER_REQ_MAP_FLAG_FUA) {
+                    if (fsync(backing_fd) < 0) {
+                        perror("Unable to fsync(), just sync()ing instead");
+                        sync();
+                    }
+                }
+                msg.type = DM_USER_RESP_SUCCESS;
+                if (lseek64(backing_fd, msg.sector * SECTOR_SIZE, SEEK_SET) < 0) {
+                    perror("Unable to seek");
+                    return -1;
+                }
+                if (not_splice(control_fd, backing_fd, msg.len) < 0) {
+                    if (errno == ENOTBLK) return 0;
+                    std::cerr << "unable to handle write data\n";
+                    return -1;
+                }
+                if (msg.flags & DM_USER_REQ_MAP_FLAG_FUA) {
+                    if (fsync(backing_fd) < 0) {
+                        perror("Unable to fsync(), just sync()ing instead");
+                        sync();
+                    }
+                }
+                break;
+            case DM_USER_REQ_MAP_FLUSH:
+                msg.type = DM_USER_RESP_SUCCESS;
+                if (fsync(backing_fd) < 0) {
+                    perror("Unable to fsync(), just sync()ing instead");
+                    sync();
+                }
+                break;
+            default:
+                std::cerr << "dmuserd: unsupported op " << std::to_string(msg.type) << "\n";
+                msg.type = DM_USER_RESP_UNSUPPORTED;
+                break;
+        }
+
+        if (verbose) std::cerr << "dmuserd: Responding to message\n";
+
+        if (write_all(control_fd, &msg, sizeof(msg)) < 0) {
+            if (errno == ENOTBLK) return 0;
+            perror("unable to write msg");
+            return -1;
+        }
+
+        switch (type) {
+            case DM_USER_REQ_MAP_READ:
+                if (verbose) std::cerr << "dmuserd: Sending read data\n";
+                if (lseek64(backing_fd, msg.sector * SECTOR_SIZE, SEEK_SET) < 0) {
+                    perror("Unable to seek");
+                    return -1;
+                }
+                if (not_splice(backing_fd, control_fd, msg.len) < 0) {
+                    if (errno == ENOTBLK) return 0;
+                    std::cerr << "unable to handle read data\n";
+                    return -1;
+                }
+                break;
+        }
+    }
+
+    /* The daemon doesn't actully terminate for this test. */
+    perror("Unable to read from control device");
+    return -1;
+}
+
+void usage(char* prog) {
+    printf("Usage: %s\n", prog);
+    printf("	Handles block requests in userspace, backed by memory\n");
+    printf("  -h			Display this help message\n");
+    printf("  -c <control dev>		Control device to use for the test\n");
+    printf("  -b <store path>		The file to use as a backing store, otherwise memory\n");
+    printf("  -v                        Enable verbose mode\n");
+}
+
+int main(int argc, char* argv[]) {
+    char* control_path = NULL;
+    char* backing_path = NULL;
+    char* store;
+    int c;
+
+    prctl(PR_SET_IO_FLUSHER, 0, 0, 0, 0);
+
+    while ((c = getopt(argc, argv, "h:c:s:b:v")) != -1) {
+        switch (c) {
+            case 'h':
+                usage(basename(argv[0]));
+                exit(0);
+            case 'c':
+                control_path = strdup(optarg);
+                break;
+            case 'b':
+                backing_path = strdup(optarg);
+                break;
+            case 'v':
+                verbose = true;
+                break;
+            default:
+                usage(basename(argv[0]));
+                exit(1);
+        }
+    }
+
+    int r = simple_daemon(control_path, backing_path);
+    if (r) fprintf(stderr, "simple_daemon() errored out\n");
+    return r;
+}
diff --git a/init/Android.mk b/init/Android.mk
index ac31ef1..4c1665b 100644
--- a/init/Android.mk
+++ b/init/Android.mk
@@ -82,6 +82,7 @@
     $(TARGET_RAMDISK_OUT)/proc \
     $(TARGET_RAMDISK_OUT)/second_stage_resources \
     $(TARGET_RAMDISK_OUT)/sys \
+    $(TARGET_RAMDISK_OUT)/metadata \
 
 LOCAL_STATIC_LIBRARIES := \
     libc++fs \
diff --git a/init/README.md b/init/README.md
index ab6a885..bcbbfbb 100644
--- a/init/README.md
+++ b/init/README.md
@@ -451,6 +451,10 @@
   exist. And it will be truncated if dst file is a normal regular file and
   already exists.
 
+`copy_per_line <src> <dst>`
+> Copies a file line by line. Similar to copy, but useful for dst is a sysfs node
+  that doesn't handle multiple lines of data.
+
 `domainname <name>`
 > Set the domain name.
 
diff --git a/init/block_dev_initializer.cpp b/init/block_dev_initializer.cpp
index 8db9793..9c2a7bb 100644
--- a/init/block_dev_initializer.cpp
+++ b/init/block_dev_initializer.cpp
@@ -40,8 +40,8 @@
     return InitMiscDevice("device-mapper");
 }
 
-bool BlockDevInitializer::InitDmUser() {
-    return InitMiscDevice("dm-user");
+bool BlockDevInitializer::InitDmUser(const std::string& name) {
+    return InitMiscDevice("dm-user!" + name);
 }
 
 bool BlockDevInitializer::InitMiscDevice(const std::string& name) {
diff --git a/init/block_dev_initializer.h b/init/block_dev_initializer.h
index b8dd3f1..79fe4ec 100644
--- a/init/block_dev_initializer.h
+++ b/init/block_dev_initializer.h
@@ -27,7 +27,7 @@
     BlockDevInitializer();
 
     bool InitDeviceMapper();
-    bool InitDmUser();
+    bool InitDmUser(const std::string& name);
     bool InitDevices(std::set<std::string> devices);
     bool InitDmDevice(const std::string& device);
 
diff --git a/init/builtins.cpp b/init/builtins.cpp
index d00d1b1..b235d2f 100644
--- a/init/builtins.cpp
+++ b/init/builtins.cpp
@@ -88,6 +88,7 @@
 
 using android::base::Basename;
 using android::base::SetProperty;
+using android::base::Split;
 using android::base::StartsWith;
 using android::base::StringPrintf;
 using android::base::unique_fd;
@@ -968,6 +969,23 @@
     return {};
 }
 
+static Result<void> do_copy_per_line(const BuiltinArguments& args) {
+    std::string file_contents;
+    if (!android::base::ReadFileToString(args[1], &file_contents, true)) {
+        return Error() << "Could not read input file '" << args[1] << "'";
+    }
+    auto lines = Split(file_contents, "\n");
+    for (const auto& line : lines) {
+        auto result = WriteFile(args[2], line);
+        if (!result.ok()) {
+            LOG(VERBOSE) << "Could not write to output file '" << args[2] << "' with '" << line
+                         << "' : " << result.error();
+        }
+    }
+
+    return {};
+}
+
 static Result<void> do_chown(const BuiltinArguments& args) {
     auto uid = DecodeUid(args[1]);
     if (!uid.ok()) {
@@ -1366,6 +1384,7 @@
         {"class_start_post_data",   {1,     1,    {false,  do_class_start_post_data}}},
         {"class_stop",              {1,     1,    {false,  do_class_stop}}},
         {"copy",                    {2,     2,    {true,   do_copy}}},
+        {"copy_per_line",           {2,     2,    {true,   do_copy_per_line}}},
         {"domainname",              {1,     1,    {true,   do_domainname}}},
         {"enable",                  {1,     1,    {false,  do_enable}}},
         {"exec",                    {1,     kMax, {false,  do_exec}}},
diff --git a/init/first_stage_init.cpp b/init/first_stage_init.cpp
index 91aaffd..83a32e7 100644
--- a/init/first_stage_init.cpp
+++ b/init/first_stage_init.cpp
@@ -117,7 +117,7 @@
 
     auto dst_dir = android::base::Dirname(dst);
     std::error_code ec;
-    if (!fs::create_directories(dst_dir, ec)) {
+    if (!fs::create_directories(dst_dir, ec) && !!ec) {
         LOG(FATAL) << "Cannot create " << dst_dir << ": " << ec.message();
     }
     if (rename(src, dst) != 0) {
@@ -315,7 +315,7 @@
         std::string dest = GetRamdiskPropForSecondStage();
         std::string dir = android::base::Dirname(dest);
         std::error_code ec;
-        if (!fs::create_directories(dir, ec)) {
+        if (!fs::create_directories(dir, ec) && !!ec) {
             LOG(FATAL) << "Can't mkdir " << dir << ": " << ec.message();
         }
         if (!fs::copy_file(kBootImageRamdiskProp, dest, ec)) {
diff --git a/init/first_stage_mount.cpp b/init/first_stage_mount.cpp
index b7d50cf..a0511cc 100644
--- a/init/first_stage_mount.cpp
+++ b/init/first_stage_mount.cpp
@@ -343,6 +343,15 @@
             if (!InitRequiredDevices({"userdata"})) {
                 return false;
             }
+            sm->SetUeventRegenCallback([this](const std::string& device) -> bool {
+                if (android::base::StartsWith(device, "/dev/block/dm-")) {
+                    return block_dev_init_.InitDmDevice(device);
+                }
+                if (android::base::StartsWith(device, "/dev/dm-user/")) {
+                    return block_dev_init_.InitDmUser(android::base::Basename(device));
+                }
+                return block_dev_init_.InitDevices({device});
+            });
             return sm->CreateLogicalAndSnapshotPartitions(super_path_);
         }
     }
diff --git a/init/host_init_verifier.cpp b/init/host_init_verifier.cpp
index ef9a451..db127d3 100644
--- a/init/host_init_verifier.cpp
+++ b/init/host_init_verifier.cpp
@@ -25,6 +25,8 @@
 #include <fstream>
 #include <iostream>
 #include <iterator>
+#include <map>
+#include <set>
 #include <string>
 #include <vector>
 
@@ -51,6 +53,7 @@
 
 using namespace std::literals;
 
+using android::base::EndsWith;
 using android::base::ParseInt;
 using android::base::ReadFileToString;
 using android::base::Split;
@@ -61,6 +64,10 @@
 
 static std::vector<std::string> passwd_files;
 
+// NOTE: Keep this in sync with the order used by init.cpp LoadBootScripts()
+static const std::vector<std::string> partition_search_order =
+        std::vector<std::string>({"system", "system_ext", "odm", "vendor", "product"});
+
 static std::vector<std::pair<std::string, int>> GetVendorPasswd(const std::string& passwd_file) {
     std::string passwd;
     if (!ReadFileToString(passwd_file, &passwd)) {
@@ -148,13 +155,24 @@
 #include "generated_stub_builtin_function_map.h"
 
 void PrintUsage() {
-    std::cout << "usage: host_init_verifier [options] <init rc file>\n"
-                 "\n"
-                 "Tests an init script for correctness\n"
-                 "\n"
-                 "-p FILE\tSearch this passwd file for users and groups\n"
-                 "--property_contexts=FILE\t Use this file for property_contexts\n"
-              << std::endl;
+    fprintf(stdout, R"(usage: host_init_verifier [options]
+
+Tests init script(s) for correctness.
+
+Generic options:
+  -p FILE                     Search this passwd file for users and groups.
+  --property_contexts=FILE    Use this file for property_contexts.
+
+Single script mode options:
+  [init rc file]              Positional argument; test this init script.
+
+Multiple script mode options:
+  --out_system=DIR            Path to the output product directory for the system partition.
+  --out_system_ext=DIR        Path to the output product directory for the system_ext partition.
+  --out_odm=DIR               Path to the output product directory for the odm partition.
+  --out_vendor=DIR            Path to the output product directory for the vendor partition.
+  --out_product=DIR           Path to the output product directory for the product partition.
+)");
 }
 
 Result<InterfaceInheritanceHierarchyMap> ReadInterfaceInheritanceHierarchy() {
@@ -203,12 +221,18 @@
     android::base::SetMinimumLogSeverity(android::base::ERROR);
 
     auto property_infos = std::vector<PropertyInfoEntry>();
+    std::map<std::string, std::string> partition_map;
 
     while (true) {
         static const char kPropertyContexts[] = "property-contexts=";
         static const struct option long_options[] = {
                 {"help", no_argument, nullptr, 'h'},
                 {kPropertyContexts, required_argument, nullptr, 0},
+                {"out_system", required_argument, nullptr, 0},
+                {"out_system_ext", required_argument, nullptr, 0},
+                {"out_odm", required_argument, nullptr, 0},
+                {"out_vendor", required_argument, nullptr, 0},
+                {"out_product", required_argument, nullptr, 0},
                 {nullptr, 0, nullptr, 0},
         };
 
@@ -224,6 +248,16 @@
                 if (long_options[option_index].name == kPropertyContexts) {
                     HandlePropertyContexts(optarg, &property_infos);
                 }
+                for (const auto& p : partition_search_order) {
+                    if (long_options[option_index].name == "out_" + p) {
+                        if (partition_map.find(p) != partition_map.end()) {
+                            PrintUsage();
+                            return EXIT_FAILURE;
+                        }
+                        partition_map[p] =
+                                EndsWith(optarg, "/") ? optarg : std::string(optarg) + "/";
+                    }
+                }
                 break;
             case 'h':
                 PrintUsage();
@@ -240,7 +274,9 @@
     argc -= optind;
     argv += optind;
 
-    if (argc != 1) {
+    // If provided, use the partition map to check multiple init rc files.
+    // Otherwise, check a single init rc file.
+    if ((!partition_map.empty() && argc != 0) || (partition_map.empty() && argc != 1)) {
         PrintUsage();
         return EXIT_FAILURE;
     }
@@ -262,24 +298,42 @@
 
     property_info_area = reinterpret_cast<const PropertyInfoArea*>(serialized_contexts.c_str());
 
+    if (!partition_map.empty()) {
+        std::vector<std::string> vendor_prefixes;
+        for (const auto& partition : {"vendor", "odm"}) {
+            if (partition_map.find(partition) != partition_map.end()) {
+                vendor_prefixes.push_back(partition_map.at(partition));
+            }
+        }
+        InitializeHostSubcontext(vendor_prefixes);
+    }
+
     const BuiltinFunctionMap& function_map = GetBuiltinFunctionMap();
     Action::set_function_map(&function_map);
     ActionManager& am = ActionManager::GetInstance();
     ServiceList& sl = ServiceList::GetInstance();
     Parser parser;
-    parser.AddSectionParser("service", std::make_unique<ServiceParser>(
-                                               &sl, nullptr, *interface_inheritance_hierarchy_map));
-    parser.AddSectionParser("on", std::make_unique<ActionParser>(&am, nullptr));
+    parser.AddSectionParser("service",
+                            std::make_unique<ServiceParser>(&sl, GetSubcontext(),
+                                                            *interface_inheritance_hierarchy_map));
+    parser.AddSectionParser("on", std::make_unique<ActionParser>(&am, GetSubcontext()));
     parser.AddSectionParser("import", std::make_unique<HostImportParser>());
 
-    if (!parser.ParseConfigFileInsecure(*argv)) {
-        LOG(ERROR) << "Failed to open init rc script '" << *argv << "'";
-        return EXIT_FAILURE;
+    if (!partition_map.empty()) {
+        for (const auto& p : partition_search_order) {
+            if (partition_map.find(p) != partition_map.end()) {
+                parser.ParseConfig(partition_map.at(p) + "etc/init");
+            }
+        }
+    } else {
+        if (!parser.ParseConfigFileInsecure(*argv)) {
+            LOG(ERROR) << "Failed to open init rc script '" << *argv << "'";
+            return EXIT_FAILURE;
+        }
     }
     size_t failures = parser.parse_error_count() + am.CheckAllCommands() + sl.CheckAllCommands();
     if (failures > 0) {
-        LOG(ERROR) << "Failed to parse init script '" << *argv << "' with " << failures
-                   << " errors";
+        LOG(ERROR) << "Failed to parse init scripts with " << failures << " error(s).";
         return EXIT_FAILURE;
     }
     return EXIT_SUCCESS;
diff --git a/init/init.cpp b/init/init.cpp
index c6f2066..1d0a9dc 100644
--- a/init/init.cpp
+++ b/init/init.cpp
@@ -875,13 +875,13 @@
     SetProperty(gsi::kGsiInstalledProp, is_installed);
 
     am.QueueBuiltinAction(SetupCgroupsAction, "SetupCgroups");
-    am.QueueBuiltinAction(TransitionSnapuserdAction, "TransitionSnapuserd");
     am.QueueBuiltinAction(SetKptrRestrictAction, "SetKptrRestrict");
     am.QueueBuiltinAction(TestPerfEventSelinuxAction, "TestPerfEventSelinux");
     am.QueueEventTrigger("early-init");
 
     // Queue an action that waits for coldboot done so we know ueventd has set up all of /dev...
     am.QueueBuiltinAction(wait_for_coldboot_done_action, "wait_for_coldboot_done");
+    am.QueueBuiltinAction(TransitionSnapuserdAction, "TransitionSnapuserd");
     // ... so that we can start queuing up actions that require stuff from /dev.
     am.QueueBuiltinAction(MixHwrngIntoLinuxRngAction, "MixHwrngIntoLinuxRng");
     am.QueueBuiltinAction(SetMmapRndBitsAction, "SetMmapRndBits");
diff --git a/init/service.cpp b/init/service.cpp
index 7b98392..766eb5d 100644
--- a/init/service.cpp
+++ b/init/service.cpp
@@ -154,6 +154,7 @@
                  .priority = 0},
       namespaces_{.flags = namespace_flags},
       seclabel_(seclabel),
+      subcontext_(subcontext_for_restart_commands),
       onrestart_(false, subcontext_for_restart_commands, "<Service '" + name + "' onrestart>", 0,
                  "onrestart", {}),
       oom_score_adjust_(DEFAULT_OOM_SCORE_ADJUST),
diff --git a/init/service.h b/init/service.h
index bc5c90f..aee1e5d 100644
--- a/init/service.h
+++ b/init/service.h
@@ -137,6 +137,7 @@
             flags_ &= ~SVC_ONESHOT;
         }
     }
+    Subcontext* subcontext() const { return subcontext_; }
 
   private:
     void NotifyStateChange(const std::string& new_state) const;
@@ -168,6 +169,7 @@
     std::vector<FileDescriptor> files_;
     std::vector<std::pair<std::string, std::string>> environment_vars_;
 
+    Subcontext* subcontext_;
     Action onrestart_;  // Commands to execute on restart.
 
     std::vector<std::string> writepid_files_;
diff --git a/init/service_parser.cpp b/init/service_parser.cpp
index 97621da..57c311a 100644
--- a/init/service_parser.cpp
+++ b/init/service_parser.cpp
@@ -657,6 +657,14 @@
                            << "' with a config in APEX";
         }
 
+        std::string context = service_->subcontext() ? service_->subcontext()->context() : "";
+        std::string old_context =
+                old_service->subcontext() ? old_service->subcontext()->context() : "";
+        if (context != old_context) {
+            return Error() << "service '" << service_->name() << "' overrides another service "
+                           << "across the treble boundary.";
+        }
+
         service_list_->RemoveService(*old_service);
         old_service = nullptr;
     }
diff --git a/init/subcontext.cpp b/init/subcontext.cpp
index dc2455e..f1fbffe 100644
--- a/init/subcontext.cpp
+++ b/init/subcontext.cpp
@@ -342,6 +342,9 @@
                 new Subcontext(std::vector<std::string>{"/vendor", "/odm"}, kVendorContext));
     }
 }
+void InitializeHostSubcontext(std::vector<std::string> vendor_prefixes) {
+    subcontext.reset(new Subcontext(vendor_prefixes, kVendorContext, /*host=*/true));
+}
 
 Subcontext* GetSubcontext() {
     return subcontext.get();
diff --git a/init/subcontext.h b/init/subcontext.h
index 788d3be..cb4138e 100644
--- a/init/subcontext.h
+++ b/init/subcontext.h
@@ -36,9 +36,11 @@
 
 class Subcontext {
   public:
-    Subcontext(std::vector<std::string> path_prefixes, std::string context)
+    Subcontext(std::vector<std::string> path_prefixes, std::string context, bool host = false)
         : path_prefixes_(std::move(path_prefixes)), context_(std::move(context)), pid_(0) {
-        Fork();
+        if (!host) {
+            Fork();
+        }
     }
 
     Result<void> Execute(const std::vector<std::string>& args);
@@ -61,6 +63,7 @@
 
 int SubcontextMain(int argc, char** argv, const BuiltinFunctionMap* function_map);
 void InitializeSubcontext();
+void InitializeHostSubcontext(std::vector<std::string> vendor_prefixes);
 Subcontext* GetSubcontext();
 bool SubcontextChildReap(pid_t pid);
 void SubcontextTerminate();
diff --git a/libcutils/fs_config.cpp b/libcutils/fs_config.cpp
index 31e1679..79c3abc 100644
--- a/libcutils/fs_config.cpp
+++ b/libcutils/fs_config.cpp
@@ -203,9 +203,14 @@
                                            CAP_MASK_LONG(CAP_SETGID),
                                               "system/bin/simpleperf_app_runner" },
     { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/e2fsck" },
-    { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/tune2fs" },
+#ifdef __LP64__
+    { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/linker64" },
+#else
+    { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/linker" },
+#endif
     { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/resize2fs" },
     { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/snapuserd" },
+    { 00755, AID_ROOT,      AID_ROOT,      0, "first_stage_ramdisk/system/bin/tune2fs" },
     // generic defaults
     { 00755, AID_ROOT,      AID_ROOT,      0, "bin/*" },
     { 00640, AID_ROOT,      AID_SHELL,     0, "fstab.*" },
diff --git a/libprocessgroup/profiles/task_profiles.json b/libprocessgroup/profiles/task_profiles.json
index ea0064f..b528fa5 100644
--- a/libprocessgroup/profiles/task_profiles.json
+++ b/libprocessgroup/profiles/task_profiles.json
@@ -100,7 +100,7 @@
           "Params":
           {
             "Controller": "cpu",
-            "Path": ""
+            "Path": "system"
           }
         }
       ]
diff --git a/libsparse/Android.bp b/libsparse/Android.bp
index bf06bbc..860b9ae 100644
--- a/libsparse/Android.bp
+++ b/libsparse/Android.bp
@@ -4,6 +4,7 @@
     name: "libsparse",
     host_supported: true,
     ramdisk_available: true,
+    vendor_ramdisk_available: true,
     recovery_available: true,
     unique_host_soname: true,
     vendor_available: true,
diff --git a/rootdir/Android.mk b/rootdir/Android.mk
index c4c8768..2bceb75 100644
--- a/rootdir/Android.mk
+++ b/rootdir/Android.mk
@@ -69,7 +69,7 @@
 
 EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS :=
 ifeq ($(CLANG_COVERAGE),true)
-  EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%20m.profraw
+  EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%p-%m.profraw
 endif
 
 # Put it here instead of in init.rc module definition,
diff --git a/rootdir/init.rc b/rootdir/init.rc
index fbb48e8..52ba921 100644
--- a/rootdir/init.rc
+++ b/rootdir/init.rc
@@ -153,21 +153,56 @@
     mkdir /dev/cpuctl/background
     mkdir /dev/cpuctl/top-app
     mkdir /dev/cpuctl/rt
+    mkdir /dev/cpuctl/system
     chown system system /dev/cpuctl
     chown system system /dev/cpuctl/foreground
     chown system system /dev/cpuctl/background
     chown system system /dev/cpuctl/top-app
     chown system system /dev/cpuctl/rt
+    chown system system /dev/cpuctl/system
     chown system system /dev/cpuctl/tasks
     chown system system /dev/cpuctl/foreground/tasks
     chown system system /dev/cpuctl/background/tasks
     chown system system /dev/cpuctl/top-app/tasks
     chown system system /dev/cpuctl/rt/tasks
+    chown system system /dev/cpuctl/system/tasks
     chmod 0664 /dev/cpuctl/tasks
     chmod 0664 /dev/cpuctl/foreground/tasks
     chmod 0664 /dev/cpuctl/background/tasks
     chmod 0664 /dev/cpuctl/top-app/tasks
     chmod 0664 /dev/cpuctl/rt/tasks
+    chmod 0664 /dev/cpuctl/system/tasks
+
+    # Create a cpu group for NNAPI HAL processes
+    mkdir /dev/cpuctl/nnapi-hal
+    chown system system /dev/cpuctl/nnapi-hal
+    chown system system /dev/cpuctl/nnapi-hal/tasks
+    chmod 0664 /dev/cpuctl/nnapi-hal/tasks
+    write /dev/cpuctl/nnapi-hal/cpu.uclamp.min 1
+    write /dev/cpuctl/nnapi-hal/cpu.uclamp.latency_sensitive 1
+
+    # Android only use global RT throttling and doesn't use CONFIG_RT_GROUP_SCHED
+    # for RT group throttling. These values here are just to make sure RT threads
+    # can be migrated to those groups. These settings can be removed once we migrate
+    # to GKI kernel.
+    write /dev/cpuctl/cpu.rt_period_us 1000000
+    write /dev/cpuctl/cpu.rt_runtime_us 950000
+    # Surfaceflinger is in FG group so giving it a bit more
+    write /dev/cpuctl/foreground/cpu.rt_runtime_us 450000
+    write /dev/cpuctl/foreground/cpu.rt_period_us 1000000
+    write /dev/cpuctl/background/cpu.rt_runtime_us 100000
+    write /dev/cpuctl/background/cpu.rt_period_us 1000000
+    write /dev/cpuctl/top-app/cpu.rt_runtime_us 100000
+    write /dev/cpuctl/top-app/cpu.rt_period_us 1000000
+    write /dev/cpuctl/rt/cpu.rt_runtime_us 100000
+    write /dev/cpuctl/rt/cpu.rt_period_us 1000000
+    write /dev/cpuctl/system/cpu.rt_runtime_us 100000
+    write /dev/cpuctl/system/cpu.rt_period_us 1000000
+    write /dev/cpuctl/nnapi-hal/cpu.rt_runtime_us 100000
+    write /dev/cpuctl/nnapi-hal/cpu.rt_period_us 1000000
+
+    # Migrate root group to system subgroup
+    copy_per_line /dev/cpuctl/tasks /dev/cpuctl/system/tasks
 
     # Create an stune group for NNAPI HAL processes
     mkdir /dev/stune/nnapi-hal
@@ -177,14 +212,6 @@
     write /dev/stune/nnapi-hal/schedtune.boost 1
     write /dev/stune/nnapi-hal/schedtune.prefer_idle 1
 
-    # cpuctl hierarchy for devices using utilclamp
-    mkdir /dev/cpuctl/nnapi-hal
-    chown system system /dev/cpuctl/nnapi-hal
-    chown system system /dev/cpuctl/nnapi-hal/tasks
-    chmod 0664 /dev/cpuctl/nnapi-hal/tasks
-    write /dev/cpuctl/nnapi-hal/cpu.uclamp.min 1
-    write /dev/cpuctl/nnapi-hal/cpu.uclamp.latency_sensitive 1
-
     # Create blkio group and apply initial settings.
     # This feature needs kernel to support it, and the
     # device's init.rc must actually set the correct values.
@@ -275,8 +302,6 @@
     write /proc/sys/vm/mmap_min_addr 32768
     write /proc/sys/net/ipv4/ping_group_range "0 2147483647"
     write /proc/sys/net/unix/max_dgram_qlen 600
-    write /proc/sys/kernel/sched_rt_runtime_us 950000
-    write /proc/sys/kernel/sched_rt_period_us 1000000
 
     # Assign reasonable ceiling values for socket rcv/snd buffers.
     # These should almost always be overridden by the target per the
@@ -298,13 +323,6 @@
     # /proc/net/fib_trie leaks interface IP addresses
     chmod 0400 /proc/net/fib_trie
 
-    # Create cgroup mount points for process groups
-    chown system system /dev/cpuctl
-    chown system system /dev/cpuctl/tasks
-    chmod 0666 /dev/cpuctl/tasks
-    write /dev/cpuctl/cpu.rt_period_us 1000000
-    write /dev/cpuctl/cpu.rt_runtime_us 950000
-
     # sets up initial cpusets for ActivityManager
     # this ensures that the cpusets are present and usable, but the device's
     # init.rc must actually set the correct cpus
@@ -808,6 +826,10 @@
     wait_for_prop apexd.status activated
     perform_apex_config
 
+    # After apexes are mounted, tell keymaster early boot has ended, so it will
+    # stop allowing use of early-boot keys
+    exec - system system -- /system/bin/vdc keymaster early-boot-ended
+
     # Special-case /data/media/obb per b/64566063
     mkdir /data/media 0770 media_rw media_rw encryption=None
     exec - media_rw media_rw -- /system/bin/chattr +F /data/media
@@ -1147,3 +1169,7 @@
 
 on property:sys.boot_completed=1 && property:sys.init.userspace_reboot.in_progress=1
   setprop sys.init.userspace_reboot.in_progress ""
+
+# Migrate tasks again in case kernel threads are created during boot
+on property:sys.boot_completed=1
+  copy_per_line /dev/cpuctl/tasks /dev/cpuctl/system/tasks
diff --git a/trusty/libtrusty/trusty.c b/trusty/libtrusty/trusty.c
index ad4d8cd..f44f8b4 100644
--- a/trusty/libtrusty/trusty.c
+++ b/trusty/libtrusty/trusty.c
@@ -29,30 +29,27 @@
 
 #include <trusty/ipc.h>
 
-int tipc_connect(const char *dev_name, const char *srv_name)
-{
-	int fd;
-	int rc;
+int tipc_connect(const char* dev_name, const char* srv_name) {
+    int fd;
+    int rc;
 
-	fd = open(dev_name, O_RDWR);
-	if (fd < 0) {
-		rc = -errno;
-		ALOGE("%s: cannot open tipc device \"%s\": %s\n",
-		      __func__, dev_name, strerror(errno));
-		return rc < 0 ? rc : -1;
-	}
+    fd = TEMP_FAILURE_RETRY(open(dev_name, O_RDWR));
+    if (fd < 0) {
+        rc = -errno;
+        ALOGE("%s: cannot open tipc device \"%s\": %s\n", __func__, dev_name, strerror(errno));
+        return rc < 0 ? rc : -1;
+    }
 
-	rc = ioctl(fd, TIPC_IOC_CONNECT, srv_name);
-	if (rc < 0) {
-		rc = -errno;
-		ALOGE("%s: can't connect to tipc service \"%s\" (err=%d)\n",
-		      __func__, srv_name, errno);
-		close(fd);
-		return rc < 0 ? rc : -1;
-	}
+    rc = TEMP_FAILURE_RETRY(ioctl(fd, TIPC_IOC_CONNECT, srv_name));
+    if (rc < 0) {
+        rc = -errno;
+        ALOGE("%s: can't connect to tipc service \"%s\" (err=%d)\n", __func__, srv_name, errno);
+        close(fd);
+        return rc < 0 ? rc : -1;
+    }
 
-	ALOGV("%s: connected to \"%s\" fd %d\n", __func__, srv_name, fd);
-	return fd;
+    ALOGV("%s: connected to \"%s\" fd %d\n", __func__, srv_name, fd);
+    return fd;
 }
 
 ssize_t tipc_send(int fd, const struct iovec* iov, int iovcnt, struct trusty_shm* shms,
@@ -63,7 +60,7 @@
     req.shm = (__u64)shms;
     req.shm_cnt = (__u64)shmcnt;
 
-    int rc = ioctl(fd, TIPC_IOC_SEND_MSG, &req);
+    int rc = TEMP_FAILURE_RETRY(ioctl(fd, TIPC_IOC_SEND_MSG, &req));
     if (rc < 0) {
         ALOGE("%s: failed to send message (err=%d)\n", __func__, rc);
     }
@@ -71,7 +68,6 @@
     return rc;
 }
 
-void tipc_close(int fd)
-{
-	close(fd);
+void tipc_close(int fd) {
+    close(fd);
 }
diff --git a/trusty/trusty-test.mk b/trusty/trusty-test.mk
index dc4c962..74106ec 100644
--- a/trusty/trusty-test.mk
+++ b/trusty/trusty-test.mk
@@ -15,4 +15,5 @@
 PRODUCT_PACKAGES += \
 	spiproxyd \
 	trusty_keymaster_set_attestation_key \
-	keymaster_soft_attestation_keys.xml \
\ No newline at end of file
+	keymaster_soft_attestation_keys.xml \
+